1/*
2 * acpi_pad.c ACPI Processor Aggregator Driver
3 *
4 * Copyright (c) 2009, Intel Corporation.
5 *
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms and conditions of the GNU General Public License,
8 * version 2, as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13 * more details.
14 *
15 */
16
17#include <linux/kernel.h>
18#include <linux/cpumask.h>
19#include <linux/module.h>
20#include <linux/init.h>
21#include <linux/types.h>
22#include <linux/kthread.h>
23#include <linux/freezer.h>
24#include <linux/cpu.h>
25#include <linux/tick.h>
26#include <linux/slab.h>
27#include <linux/acpi.h>
28#include <asm/mwait.h>
29
30#define ACPI_PROCESSOR_AGGREGATOR_CLASS	"acpi_pad"
31#define ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME "Processor Aggregator"
32#define ACPI_PROCESSOR_AGGREGATOR_NOTIFY 0x80
33static DEFINE_MUTEX(isolated_cpus_lock);
34static DEFINE_MUTEX(round_robin_lock);
35
36static unsigned long power_saving_mwait_eax;
37
38static unsigned char tsc_detected_unstable;
39static unsigned char tsc_marked_unstable;
40
41static void power_saving_mwait_init(void)
42{
43	unsigned int eax, ebx, ecx, edx;
44	unsigned int highest_cstate = 0;
45	unsigned int highest_subcstate = 0;
46	int i;
47
48	if (!boot_cpu_has(X86_FEATURE_MWAIT))
49		return;
50	if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
51		return;
52
53	cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
54
55	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
56	    !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
57		return;
58
59	edx >>= MWAIT_SUBSTATE_SIZE;
60	for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
61		if (edx & MWAIT_SUBSTATE_MASK) {
62			highest_cstate = i;
63			highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
64		}
65	}
66	power_saving_mwait_eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
67		(highest_subcstate - 1);
68
69#if defined(CONFIG_X86)
70	switch (boot_cpu_data.x86_vendor) {
71	case X86_VENDOR_AMD:
72	case X86_VENDOR_INTEL:
73		/*
74		 * AMD Fam10h TSC will tick in all
75		 * C/P/S0/S1 states when this bit is set.
76		 */
77		if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
78			tsc_detected_unstable = 1;
79		break;
80	default:
81		/* TSC could halt in idle */
82		tsc_detected_unstable = 1;
83	}
84#endif
85}
86
87static unsigned long cpu_weight[NR_CPUS];
88static int tsk_in_cpu[NR_CPUS] = {[0 ... NR_CPUS-1] = -1};
89static DECLARE_BITMAP(pad_busy_cpus_bits, NR_CPUS);
90static void round_robin_cpu(unsigned int tsk_index)
91{
92	struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
93	cpumask_var_t tmp;
94	int cpu;
95	unsigned long min_weight = -1;
96	unsigned long uninitialized_var(preferred_cpu);
97
98	if (!alloc_cpumask_var(&tmp, GFP_KERNEL))
99		return;
100
101	mutex_lock(&round_robin_lock);
102	cpumask_clear(tmp);
103	for_each_cpu(cpu, pad_busy_cpus)
104		cpumask_or(tmp, tmp, topology_sibling_cpumask(cpu));
105	cpumask_andnot(tmp, cpu_online_mask, tmp);
106	/* avoid HT sibilings if possible */
107	if (cpumask_empty(tmp))
108		cpumask_andnot(tmp, cpu_online_mask, pad_busy_cpus);
109	if (cpumask_empty(tmp)) {
110		mutex_unlock(&round_robin_lock);
111		return;
112	}
113	for_each_cpu(cpu, tmp) {
114		if (cpu_weight[cpu] < min_weight) {
115			min_weight = cpu_weight[cpu];
116			preferred_cpu = cpu;
117		}
118	}
119
120	if (tsk_in_cpu[tsk_index] != -1)
121		cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
122	tsk_in_cpu[tsk_index] = preferred_cpu;
123	cpumask_set_cpu(preferred_cpu, pad_busy_cpus);
124	cpu_weight[preferred_cpu]++;
125	mutex_unlock(&round_robin_lock);
126
127	set_cpus_allowed_ptr(current, cpumask_of(preferred_cpu));
128}
129
130static void exit_round_robin(unsigned int tsk_index)
131{
132	struct cpumask *pad_busy_cpus = to_cpumask(pad_busy_cpus_bits);
133	cpumask_clear_cpu(tsk_in_cpu[tsk_index], pad_busy_cpus);
134	tsk_in_cpu[tsk_index] = -1;
135}
136
137static unsigned int idle_pct = 5; /* percentage */
138static unsigned int round_robin_time = 1; /* second */
139static int power_saving_thread(void *data)
140{
141	struct sched_param param = {.sched_priority = 1};
142	int do_sleep;
143	unsigned int tsk_index = (unsigned long)data;
144	u64 last_jiffies = 0;
145
146	sched_setscheduler(current, SCHED_RR, &param);
147
148	while (!kthread_should_stop()) {
149		unsigned long expire_time;
150
151		/* round robin to cpus */
152		expire_time = last_jiffies + round_robin_time * HZ;
153		if (time_before(expire_time, jiffies)) {
154			last_jiffies = jiffies;
155			round_robin_cpu(tsk_index);
156		}
157
158		do_sleep = 0;
159
160		expire_time = jiffies + HZ * (100 - idle_pct) / 100;
161
162		while (!need_resched()) {
163			if (tsc_detected_unstable && !tsc_marked_unstable) {
164				/* TSC could halt in idle, so notify users */
165				mark_tsc_unstable("TSC halts in idle");
166				tsc_marked_unstable = 1;
167			}
168			local_irq_disable();
169			tick_broadcast_enable();
170			tick_broadcast_enter();
171			stop_critical_timings();
172
173			mwait_idle_with_hints(power_saving_mwait_eax, 1);
174
175			start_critical_timings();
176			tick_broadcast_exit();
177			local_irq_enable();
178
179			if (time_before(expire_time, jiffies)) {
180				do_sleep = 1;
181				break;
182			}
183		}
184
185		/*
186		 * current sched_rt has threshold for rt task running time.
187		 * When a rt task uses 95% CPU time, the rt thread will be
188		 * scheduled out for 5% CPU time to not starve other tasks. But
189		 * the mechanism only works when all CPUs have RT task running,
190		 * as if one CPU hasn't RT task, RT task from other CPUs will
191		 * borrow CPU time from this CPU and cause RT task use > 95%
192		 * CPU time. To make 'avoid starvation' work, takes a nap here.
193		 */
194		if (unlikely(do_sleep))
195			schedule_timeout_killable(HZ * idle_pct / 100);
196
197		/* If an external event has set the need_resched flag, then
198		 * we need to deal with it, or this loop will continue to
199		 * spin without calling __mwait().
200		 */
201		if (unlikely(need_resched()))
202			schedule();
203	}
204
205	exit_round_robin(tsk_index);
206	return 0;
207}
208
209static struct task_struct *ps_tsks[NR_CPUS];
210static unsigned int ps_tsk_num;
211static int create_power_saving_task(void)
212{
213	int rc;
214
215	ps_tsks[ps_tsk_num] = kthread_run(power_saving_thread,
216		(void *)(unsigned long)ps_tsk_num,
217		"acpi_pad/%d", ps_tsk_num);
218
219	if (IS_ERR(ps_tsks[ps_tsk_num])) {
220		rc = PTR_ERR(ps_tsks[ps_tsk_num]);
221		ps_tsks[ps_tsk_num] = NULL;
222	} else {
223		rc = 0;
224		ps_tsk_num++;
225	}
226
227	return rc;
228}
229
230static void destroy_power_saving_task(void)
231{
232	if (ps_tsk_num > 0) {
233		ps_tsk_num--;
234		kthread_stop(ps_tsks[ps_tsk_num]);
235		ps_tsks[ps_tsk_num] = NULL;
236	}
237}
238
239static void set_power_saving_task_num(unsigned int num)
240{
241	if (num > ps_tsk_num) {
242		while (ps_tsk_num < num) {
243			if (create_power_saving_task())
244				return;
245		}
246	} else if (num < ps_tsk_num) {
247		while (ps_tsk_num > num)
248			destroy_power_saving_task();
249	}
250}
251
252static void acpi_pad_idle_cpus(unsigned int num_cpus)
253{
254	get_online_cpus();
255
256	num_cpus = min_t(unsigned int, num_cpus, num_online_cpus());
257	set_power_saving_task_num(num_cpus);
258
259	put_online_cpus();
260}
261
262static uint32_t acpi_pad_idle_cpus_num(void)
263{
264	return ps_tsk_num;
265}
266
267static ssize_t acpi_pad_rrtime_store(struct device *dev,
268	struct device_attribute *attr, const char *buf, size_t count)
269{
270	unsigned long num;
271	if (kstrtoul(buf, 0, &num))
272		return -EINVAL;
273	if (num < 1 || num >= 100)
274		return -EINVAL;
275	mutex_lock(&isolated_cpus_lock);
276	round_robin_time = num;
277	mutex_unlock(&isolated_cpus_lock);
278	return count;
279}
280
281static ssize_t acpi_pad_rrtime_show(struct device *dev,
282	struct device_attribute *attr, char *buf)
283{
284	return scnprintf(buf, PAGE_SIZE, "%d\n", round_robin_time);
285}
286static DEVICE_ATTR(rrtime, S_IRUGO|S_IWUSR,
287	acpi_pad_rrtime_show,
288	acpi_pad_rrtime_store);
289
290static ssize_t acpi_pad_idlepct_store(struct device *dev,
291	struct device_attribute *attr, const char *buf, size_t count)
292{
293	unsigned long num;
294	if (kstrtoul(buf, 0, &num))
295		return -EINVAL;
296	if (num < 1 || num >= 100)
297		return -EINVAL;
298	mutex_lock(&isolated_cpus_lock);
299	idle_pct = num;
300	mutex_unlock(&isolated_cpus_lock);
301	return count;
302}
303
304static ssize_t acpi_pad_idlepct_show(struct device *dev,
305	struct device_attribute *attr, char *buf)
306{
307	return scnprintf(buf, PAGE_SIZE, "%d\n", idle_pct);
308}
309static DEVICE_ATTR(idlepct, S_IRUGO|S_IWUSR,
310	acpi_pad_idlepct_show,
311	acpi_pad_idlepct_store);
312
313static ssize_t acpi_pad_idlecpus_store(struct device *dev,
314	struct device_attribute *attr, const char *buf, size_t count)
315{
316	unsigned long num;
317	if (kstrtoul(buf, 0, &num))
318		return -EINVAL;
319	mutex_lock(&isolated_cpus_lock);
320	acpi_pad_idle_cpus(num);
321	mutex_unlock(&isolated_cpus_lock);
322	return count;
323}
324
325static ssize_t acpi_pad_idlecpus_show(struct device *dev,
326	struct device_attribute *attr, char *buf)
327{
328	return cpumap_print_to_pagebuf(false, buf,
329				       to_cpumask(pad_busy_cpus_bits));
330}
331
332static DEVICE_ATTR(idlecpus, S_IRUGO|S_IWUSR,
333	acpi_pad_idlecpus_show,
334	acpi_pad_idlecpus_store);
335
336static int acpi_pad_add_sysfs(struct acpi_device *device)
337{
338	int result;
339
340	result = device_create_file(&device->dev, &dev_attr_idlecpus);
341	if (result)
342		return -ENODEV;
343	result = device_create_file(&device->dev, &dev_attr_idlepct);
344	if (result) {
345		device_remove_file(&device->dev, &dev_attr_idlecpus);
346		return -ENODEV;
347	}
348	result = device_create_file(&device->dev, &dev_attr_rrtime);
349	if (result) {
350		device_remove_file(&device->dev, &dev_attr_idlecpus);
351		device_remove_file(&device->dev, &dev_attr_idlepct);
352		return -ENODEV;
353	}
354	return 0;
355}
356
357static void acpi_pad_remove_sysfs(struct acpi_device *device)
358{
359	device_remove_file(&device->dev, &dev_attr_idlecpus);
360	device_remove_file(&device->dev, &dev_attr_idlepct);
361	device_remove_file(&device->dev, &dev_attr_rrtime);
362}
363
364/*
365 * Query firmware how many CPUs should be idle
366 * return -1 on failure
367 */
368static int acpi_pad_pur(acpi_handle handle)
369{
370	struct acpi_buffer buffer = {ACPI_ALLOCATE_BUFFER, NULL};
371	union acpi_object *package;
372	int num = -1;
373
374	if (ACPI_FAILURE(acpi_evaluate_object(handle, "_PUR", NULL, &buffer)))
375		return num;
376
377	if (!buffer.length || !buffer.pointer)
378		return num;
379
380	package = buffer.pointer;
381
382	if (package->type == ACPI_TYPE_PACKAGE &&
383		package->package.count == 2 &&
384		package->package.elements[0].integer.value == 1) /* rev 1 */
385
386		num = package->package.elements[1].integer.value;
387
388	kfree(buffer.pointer);
389	return num;
390}
391
392static void acpi_pad_handle_notify(acpi_handle handle)
393{
394	int num_cpus;
395	uint32_t idle_cpus;
396	struct acpi_buffer param = {
397		.length = 4,
398		.pointer = (void *)&idle_cpus,
399	};
400
401	mutex_lock(&isolated_cpus_lock);
402	num_cpus = acpi_pad_pur(handle);
403	if (num_cpus < 0) {
404		mutex_unlock(&isolated_cpus_lock);
405		return;
406	}
407	acpi_pad_idle_cpus(num_cpus);
408	idle_cpus = acpi_pad_idle_cpus_num();
409	acpi_evaluate_ost(handle, ACPI_PROCESSOR_AGGREGATOR_NOTIFY, 0, &param);
410	mutex_unlock(&isolated_cpus_lock);
411}
412
413static void acpi_pad_notify(acpi_handle handle, u32 event,
414	void *data)
415{
416	struct acpi_device *device = data;
417
418	switch (event) {
419	case ACPI_PROCESSOR_AGGREGATOR_NOTIFY:
420		acpi_pad_handle_notify(handle);
421		acpi_bus_generate_netlink_event(device->pnp.device_class,
422			dev_name(&device->dev), event, 0);
423		break;
424	default:
425		pr_warn("Unsupported event [0x%x]\n", event);
426		break;
427	}
428}
429
430static int acpi_pad_add(struct acpi_device *device)
431{
432	acpi_status status;
433
434	strcpy(acpi_device_name(device), ACPI_PROCESSOR_AGGREGATOR_DEVICE_NAME);
435	strcpy(acpi_device_class(device), ACPI_PROCESSOR_AGGREGATOR_CLASS);
436
437	if (acpi_pad_add_sysfs(device))
438		return -ENODEV;
439
440	status = acpi_install_notify_handler(device->handle,
441		ACPI_DEVICE_NOTIFY, acpi_pad_notify, device);
442	if (ACPI_FAILURE(status)) {
443		acpi_pad_remove_sysfs(device);
444		return -ENODEV;
445	}
446
447	return 0;
448}
449
450static int acpi_pad_remove(struct acpi_device *device)
451{
452	mutex_lock(&isolated_cpus_lock);
453	acpi_pad_idle_cpus(0);
454	mutex_unlock(&isolated_cpus_lock);
455
456	acpi_remove_notify_handler(device->handle,
457		ACPI_DEVICE_NOTIFY, acpi_pad_notify);
458	acpi_pad_remove_sysfs(device);
459	return 0;
460}
461
462static const struct acpi_device_id pad_device_ids[] = {
463	{"ACPI000C", 0},
464	{"", 0},
465};
466MODULE_DEVICE_TABLE(acpi, pad_device_ids);
467
468static struct acpi_driver acpi_pad_driver = {
469	.name = "processor_aggregator",
470	.class = ACPI_PROCESSOR_AGGREGATOR_CLASS,
471	.ids = pad_device_ids,
472	.ops = {
473		.add = acpi_pad_add,
474		.remove = acpi_pad_remove,
475	},
476};
477
478static int __init acpi_pad_init(void)
479{
480	power_saving_mwait_init();
481	if (power_saving_mwait_eax == 0)
482		return -EINVAL;
483
484	return acpi_bus_register_driver(&acpi_pad_driver);
485}
486
487static void __exit acpi_pad_exit(void)
488{
489	acpi_bus_unregister_driver(&acpi_pad_driver);
490}
491
492module_init(acpi_pad_init);
493module_exit(acpi_pad_exit);
494MODULE_AUTHOR("Shaohua Li<shaohua.li@intel.com>");
495MODULE_DESCRIPTION("ACPI Processor Aggregator Driver");
496MODULE_LICENSE("GPL");
497